Universal perceptron and DNA-like learning algorithm for binary neural networks: non-LSBF implementation.

Area of Science:

• Computer Science
• Boolean Algebra
• Machine Learning

Background:

• Implementing linearly nonseparable Boolean functions (non-LSBF) is challenging due to their high complexity.
• The proportion of non-LSBF increases exponentially with the number of input variables.
• Existing methods struggle with the computational demands of non-LSBF.

Purpose of the Study:

• To propose an effective algorithm for implementing non-LSBF.
• To address the complexity and computational challenges associated with non-LSBF.
• To demonstrate the algorithm's capability in function approximation and Boolean function mapping.

Main Methods:

• Introduction of the DNA-like learning and decomposing algorithm (DNA-like LDA).
• Training a DNA-like offset sequence for function decomposition.
• Decomposing non-LSBF into logic XOR operations of linearly separable Boolean functions (LSBF).
• Determining weight-threshold values for a multilayer perceptron (MLP) to perform decomposition and mapping.

Main Results:

• The DNA-like LDA successfully decomposes non-LSBF into a sequence of LSBF.
• The MLP, guided by the algorithm, effectively performs both LSBF decomposition and hidden-to-output neuron mapping.
• Validation through approximation of a circular region and the n-bit parity Boolean function (PBF) confirms algorithm efficacy.

Conclusions:

• The proposed DNA-like LDA offers an effective solution for implementing complex non-LSBF.
• This approach simplifies the implementation of non-LSBF by leveraging LSBF and MLP.
• The algorithm shows promise for applications requiring the processing of complex Boolean functions.